Artificial Intelligence in Thyroid Imaging: A Review of Deep Learning Techniques and Clinical Applications

Abstract

Thyroid diseases, encompassing both benign nodules and differentiated thyroid cancers, are highly prevalent worldwide and necessitate accurate diagnostic imaging for effective management and treatment. The traditional interpretation of medical images, however, remains subjective and is heavily reliant on the expertise of clinicians, which can lead to variability in diagnostic outcomes and treatment decisions. Recent advancements in deep learning (DL) techniques have shown considerable promise in addressing these limitations. This review aims to summarize the application of DL in thyroid disease diagnosis across multiple imaging modalities, including single-photon emission computed tomography (SPECT), ultrasound, and computed tomography (CT). We focus on five recent studies that utilized state-of-the-art DL architectures, such as residual networks (ResNet), Xception-based multi-channel models, and ensemble learning approaches, which have demonstrated remarkable efficacy in image classification and disease characterization. These DL models were evaluated based on diagnostic accuracy, clinical integration, interpretability, and real-world performance, with a direct comparison to radiologists and fine-needle aspiration (FNA). Across all imaging modalities, DL models consistently demonstrated robust performance, frequently matching or exceeding the diagnostic accuracy of experienced clinicians. For instance, diagnostic accuracies up to 98.9% were achieved in ultrasound and CT-based multi-channel models, with areas under the receiver operating characteristic curve (AUC) surpassing 0.93 in several studies. Additionally, prospective validation and the incorporation of clinical risk factors, such as patient demographics and prior medical history, further enhanced the reliability and clinical relevance of these models. Deep learning has demonstrated significant promise in enhancing thyroid imaging diagnostics by providing quicker, more standardized, and possibly noninvasive substitutes for conventional diagnostic techniques. Despite challenges related to generalizability and model interpretability, the integration of AI into thyroid imaging holds significant promise for improving diagnostic precision and enhancing the efficiency of endocrine care. As these technologies continue to evolve, they may lead to a paradigm shift in the way thyroid disorders are diagnosed and managed, moving toward a more accurate, reproducible, and patient-centered approach.

Ähnliche Arbeiten